Evaluations & Metrics

Monitor performance, optimize costs, and evaluate LLM effectiveness

Evaluations help you understand how well your automation performs, which models work best for your use cases, and how to optimize for cost and reliability. This guide covers both monitoring your own workflows and running comprehensive evaluations.

Why Evaluations Matter

• Performance Optimization: Identify which models and settings work best for your specific automation tasks
• Cost Control: Track token usage and inference time to optimize spending
• Reliability: Measure success rates and identify failure patterns
• Model Selection: Compare different LLMs on real-world tasks to make informed decisions

Live Model Comparisons

View real-time performance comparisons across different LLMs on the Stagehand Evals Dashboard

Comprehensive Evaluations

Evaluations help you systematically test and improve your automation workflows. Stagehand provides both built-in evaluations and tools to create your own.

Evals CLI

To run evals, you’ll need to clone the Stagehand repo and set up the CLI.

We recommend using Braintrust to help visualize evals results and metrics.

The Stagehand CLI provides a powerful interface for running evaluations. You can run specific evals, categories, or external benchmarks with customizable settings.

Evals are grouped into:

1. Act Evals - These are evals that test the functionality of the act method.
2. Extract Evals - These are evals that test the functionality of the extract method.
3. Observe Evals - These are evals that test the functionality of the observe method.
4. Combination Evals - These are evals that test the functionality of the act, extract, and observe methods together.
5. Experimental Evals - These are experimental custom evals that test the functionality of the stagehand primitives.
6. Agent Evals - These are evals that test the functionality of agent.
7. (NEW) External Benchmarks - Run external benchmarks like WebBench, GAIA, WebVoyager, OnlineMind2Web, and OSWorld.

Installation

1. Install Dependencies

   # From the stagehand root directory
   pnpm install
   
2. Build the CLI

   pnpm run build:cli
   
3. Verify Installation

   evals help
   

CLI Commands and Options

Basic Commands

# Run all evals
evals run all

# Run specific category
evals run act
evals run extract
evals run observe
evals run agent

# Run specific eval
evals run extract/extract_text

# List available evals
evals list
evals list --detailed

# Configure defaults
evals config
evals config set env browserbase
evals config set trials 5

Command Options

• -e, --env: Environment (local or browserbase)
• -t, --trials: Number of trials per eval (default: 3)
• -c, --concurrency: Max parallel sessions (default: 10)
• -m, --model: Model override
• -p, --provider: Provider override
• --api: Use Stagehand API instead of SDK

Running External Benchmarks

The CLI supports several industry-standard benchmarks:

# WebBench with filters
evals run benchmark:webbench -l 10 -f difficulty=easy -f category=READ

# GAIA benchmark
evals run b:gaia -s 100 -l 25 -f level=1

# WebVoyager
evals run b:webvoyager -l 50

# OnlineMind2Web
evals run b:onlineMind2Web

# OSWorld
evals run b:osworld -f source=Mind2Web

Configuration Files

You can view the specific evals in evals/tasks. Each eval is grouped into eval categories based on evals/evals.config.json.

Viewing eval results

Eval results are viewable on Braintrust. You can view the results of a specific eval by going to the Braintrust URL specified in the terminal when you run npm run evals.

By default, each eval will run five times per model. The “Exact Match” column shows the percentage of times the eval was correct. The “Error Rate” column shows the percentage of times the eval errored out.

You can use the Braintrust UI to filter by model/eval and aggregate results across all evals.

Creating Custom Evaluations

Step-by-Step Guide

1. Create Evaluation File

   Create a new file in evals/tasks/your-eval.ts:

   import { EvalTask } from '../types';
   
   export const customEvalTask: EvalTask = {
     name: 'custom_task_name',
     description: 'Test specific automation workflow',
     
     // Test setup
     setup: async ({ page }) => {
       await page.goto('https://example.com');
     },
     
     // The actual test
     task: async ({ stagehand, page }) => {
       // Your automation logic
       await stagehand.act({ action: 'click the login button' });
       const result = await stagehand.extract({ 
         instruction: 'Get the user name',
         schema: { username: 'string' }
       });
       return result;
     },
     
     // Validation
     validate: (result, expected) => {
       return result.username === expected.username;
     },
     
     // Test cases
     testCases: [
       {
         input: { /* test input */ },
         expected: { username: 'john_doe' }
       }
     ],
     
     // Evaluation criteria
     scoring: {
       exactMatch: true,
       timeout: 30000,
       retries: 2
     }
   };
   
2. Add to Configuration

   Update evals/evals.config.json:

   {
     "categories": {
       "custom": ["custom_task_name"],
       "existing_category": ["custom_task_name"]
     }
   }
   
3. Run Your Evaluation

   # Test your custom evaluation
   evals run custom_task_name
   
   # Run the entire custom category
   evals run custom
   
   # Run with specific settings
   evals run custom_task_name -e browserbase -t 5 -m gpt-4o
   

Best Practices for Custom Evals

Test Design Principles

• Atomic: Each test should validate one specific capability
• Deterministic: Tests should produce consistent, measurable results
• Realistic: Use real-world scenarios and websites
• Measurable: Define clear success/failure criteria

Performance Optimization

• Parallel Execution: Design tests to run independently
• Resource Management: Clean up after each test
• Timeout Handling: Set appropriate timeouts for operations
• Error Recovery: Handle failures gracefully

Data Quality

• Ground Truth: Establish reliable expected outcomes
• Edge Cases: Test boundary conditions and error scenarios
• Statistical Significance: Run multiple iterations for reliability
• Version Control: Track changes to test cases over time

Troubleshooting Evaluations

Evaluation Timeouts

Symptoms: Tests fail with timeout errors

Solutions:

• Increase timeout in taskConfig.ts
• Use faster models (Gemini 2.5 Flash, Claude Haiku 4.5)
• Optimize test scenarios to be less complex
• Check network connectivity to LLM providers

Inconsistent Results

Symptoms: Same test passes/fails randomly

Solutions:

• Set temperature to 0 for deterministic outputs
• Increase repetitions for statistical significance
• Use more capable models for complex tasks
• Check for dynamic website content affecting tests

High Evaluation Costs

Symptoms: Token usage exceeding budget

Solutions:

• Use cost-effective models (Gemini 2.5 Flash, Claude Haiku 4.5)
• Reduce repetitions for initial testing
• Focus on specific evaluation categories
• Use local browser environment to reduce Browserbase costs

Braintrust Integration Issues

Symptoms: Results not uploading to dashboard

Solutions:

• Check Braintrust API key configuration
• Verify internet connectivity
• Update Braintrust SDK to latest version
• Check project permissions in Braintrust dashboard